Editmamei ships no model of its own. It runs on whatever AI client you point at it, so every edit spends that client’s tokens. I had never measured how many, so in late August I did.

The setup

Twenty three raw files from a morning at Owasco Lake, shot on a phone. Flat light, a hazy sky, my dog on a dock. I picked one frame, edited it until I liked it, and saved the look as a template. Then I ran the same work eight different ways and recorded what each one spent.

Every run drove Photoshop.

What one photo costs

Editing a single photo from a plain-language description, with no template, came to roughly 229,000 tokens and took about twenty three minutes. That is the number for starting cold: the model has to look at the picture, decide on a treatment, try things, check its own work, and change its mind a few times.

The money is not really the point. At pay-as-you-go API rates a run like that is on the order of a dollar, depending on which model you use. What you feel on a subscription is different, because a subscription meters you in rate limits rather than dollars, and a twenty-three minute agent loop is a real bite out of one. That is the form most people will run into.

Applying a saved template to one photo instead came to about 152,000 tokens in seven and a half minutes. A third cheaper, and a third of the time, because the deciding has already been done.

What a shoot costs

This is the part I did not expect.

Running the same template across ten photos cost 226,000 tokens in total. Not per photo. In total, for all ten. That works out to about 22,600 tokens each, roughly a tenth of what the first cold edit cost.

The saving is not linear, and it gets better as the pile gets bigger. Going from one photo to three, each additional photo added around 32,000 tokens. Going from three to ten, each additional photo added around 1,500. Once the look is settled and the rhythm is established, more photos are close to free.

The last step is different

Cropping the finished set for social is mechanical work, and Photoshop has its own batch engine for exactly that. Editmamei can drive it.

At three files, using the batch engine was slightly worse than just doing them one at a time. It has setup costs, and three files is not enough to pay them off. At ten files it flipped: 15% cheaper and 2.7 times faster, and the cost per extra file dropped to almost nothing while the one-at-a-time approach kept climbing.

I mention the case where it lost because I nearly published the version where I only tested three files and concluded the batch engine was not worth much. It is worth a great deal, at the size it was built for.

Where it was going

Across every run, 96% of everything the tool sent back was preview images. Ninety six. Not the edits, not the histograms, not the layer data. The pictures the model looks at to check its own work.

That matched what the timing already said: checking the work accounts for most of the wall clock. Looking is more expensive than editing, in both senses.

So I changed it. Since 1.4.0, a result only carries a preview image when you ask for one, and in a run of several steps only the last step keeps an inline preview. Every number above was measured before that shipped, which makes them an upper bound. I have not re-measured yet.

If you are about to try it

The short version: the first edit is the expensive one, and it is expensive because the model is deciding. Every edit after that, on the same look, is cheap.

So do not start by pointing it at forty photos. Edit one frame until it is right, save it as a template, and then carry it across the rest. That is the order that makes the numbers above work in your favour, and it is the order that produces a consistent set across a shoot, which is the better reason to do it.

Templates and batch are Pro features. The cold edit — describe what you want, get layers back — is the free edition.

The caveat

I ran each of these once. That is enough to see a tenfold difference and a crossover point, and not enough to publish a precise figure with confidence, because these runs vary. Treat the shapes as real and the exact numbers as a first reading. I will run repeats before I put any of this on a pricing page.

If you have edited a photo with Editmamei, I want to know what your first one cost you — whether 229,000 is normal or whether I picked an unlucky frame. Tell me.

— Alex